Abstract:

Using the case study of a sterile-type chilling disaster during the head-flowering phase of rice, this study presents a novel approach to constructing vulnerability curves that can overcome data limitations while also taking into account crop growth mechanisms. Meteorological data during 1990-2010 was used to generate sterile-type chilling scenarios at county scale, estimated rice yield losses through combining one crop model (MCWLA) and machine learning (XGBoost) method, finally developed sterile-type chilling vulnerability curves for each main rice-planting zone in China and estimated long-term historical (1961-2010) yield loss caused by sterile-type chilling disasters. The results showed that: (1) Machine learning could effectively reproduce the estimation ability of crop model (RRMSE<6%, R²>0.93). (2) The sterile-type vulnerability decreased with decreasing latitude, and was weaker in growing seasons for late rice than that in early rice. (3) The historical yield loss was higher for single rice (1224kg·ha⁻¹) than for double rice (early rice: 868kg·ha⁻¹; late rice: 807kg·ha⁻¹). 

Key words: Rice, Chilling disaster, Vulnerability, Crop model, Machine learning